“Pre-training currently constitutes the majority of the cost for developing large language models.”